Update #2: I've just gotten word from Replicate that they have re-opened the $10 link! It should be working again. :) Update #1: Apparently the link expired for Replicate. I was not told that the link would expire. I apologize for this. I honestly had no idea. I reached out to them and asked what happened and they told me that their finance team decided to turn it off because too many people were using it. I'm going to do what I can to rectify the situation, even if it means finding a different method people can follow to do it for free. Once again, I sincerely apologize. This decision was made my Replicate and they did not inform me that they would turn the link off if it got too much traffic.
thanks for your dedication, a lot of can't afford to pay for gpu time and would really appreciate a free alternative to train flux lora, maybe kaggle perhaps
@@luizpaes7307 You can generate some pretty nice stuff with Stable Diffusion 1.5 on a 2060 with the right Checkpoint/Lora but yeah, not gonna work with Flux.
Hey Matt. We followed your instructions and generated some impressive images. Thanks for the detailed practical way you do the demos. Your dedication and passion are a cut above many.
Ok, there were a couple of things that I've learned from other vids that may have gotten missed with this one. 1. It must be set to public (which I think he was later addressed) and the other way I've learned and now do with no issues is 2. When creating your TOKEN you have three options - Fine-grained, Read, Write - you want to set it to WRITE and that solved all my problems. I hope it solves yours. Good Luck.
Thank you so much! The error was driving me crazy. Unfortunately the results look nothing like me. It's probably my fault with the images I chose? There were I think 50. Maybe I should have prioritized the best ones.
Regarding your question of positioning of the trigger word: LLMs in general pay more attention to the beginning and end of a prompt than they do to the middle (something which human brains *also* do, by the way 😁). So moving the trigger word to the beginning or end will indeed have a measurable effect on how often it includes the trained face/concept correctly.
For anyone getting the error Command '['python', 'run.py', 'config/replicate.yml']' returned non-zero exit status 1: I also got the error, but it worked on the 2nd try with a couple of changes. Not sure what exactly was causing it but you can do all of these which will hopefully work for you. 1) Specify the model name instead of leaving it blank (black-forest-labs/FLUX.1-dev) 2) Rename all of your training images to be shorter (ex: person, instead of photo_of_person) (Edit: Probably not needed it's just something I randomly tried) 3) Make sure you request access to the black-forest-labs/FLUX.1-dev model on Hugging Face (I already did this prior, so probably not the issue, but who knows) 4) When creating your access token, scroll down to "Repositories permissions", select the location for your Lora, grant all permissions
It is likely step 4 that was the issue and what fixed it. I would imagine first creating the repository and then the access token would also fix the issue.
I still got an error after doing all that, then I resized my images and there were two corrupt photos that I had to delete from my zip after taking claude's help. Now it is working fine.
OMG Matt. This training is blowing my mind! I'm so excited with the outputs I'm getting. And would have been crawling into a deep dark hole of despair trying to figure this out on my own. LOVE IT!
Hey Matt, I just wanted to drop you a quick note to say how awesome I think your tutorials are. I’ve gone through the steps, and even though I haven’t had the chance to test yet, I can already tell that this is going to be a game changer for me. It’s looking like we’re getting some rainy weather next week, so that’s when I’ll really dive into it. Just wanted to say thanks for making everything so easy to follow. You’re doing an amazing job, and I’m excited to see where this takes me./ Anna
There seems to be a giant issue in the script of the ai-toolkit that causes it to fail every time now. It seems to be downloading EVERY image ever supplied to it succesfully. When you check the logs you can clearly see that it is downloading multiple sets of images with names of other people that (obviously) can't be found anymore or are so big in total size that the memory can't save them all locally anymore. Seems like a 'little' oversight on the dev's part 😂
BRO!!! thank you ! your editing amazing love the new glitch style transition, and information presentation / lesson / 10 free , wow, this is one of your best videos yet. you the man !!
2 месяца назад+9
It worked like a charm.... but I want to use the model I created on hugging face on my local ComfyUI. Can you make a short tutorial on how to do this please!! Love your videos... keep up the good work!!
You download the safetensor from your Hugging face and include it in your lora folders. Onwards, you use it as any other lora. I didn't try yet, but that is what I will do when I will create mine.
A question - what happens if you don't want to upload the model to Hugging Face at all? Do you get a download link, or something else. I read about this toolkit just yesterday but with the VRAM requirements it was a no go for me. Didn't know about Replicate but it has me interested and seems easy enough. Cheers.
Sir, for free credits it shows 'It looks like that invite has expired. Please ask whoever sent it to you to send a new one.'. Can you please send the new invite link. Thank you
Wow, thank you very much for providing the credits and your continuing great tutorials, really highly appreciated. I truly enjoy your content and am always thrilled when you upload something new. I especially love that you cover so many different AI topics and that you always take the time to teach your findings in a easy to follow way. Keep up the great work and thanks again for sharing.
Question, the $10 coupon no longer works... is there a page where I can upload or use my replicate lora? That is, use it and continue generating images on another page? (I have downloaded the lora replicate in a zip file).
I followed your instructions but after clicking an "create training" it always says: Command '['python', 'run.py', 'config/replicate.yml']' returned non-zero exit status 1. Cany ou help?
Please keep breaking the tools down like this I am taking notes here. Your tutorials have been invaluable and a new wave of awesomeness is clearly upon us.
Super appreciate this. I followed your tutorial from last year, and these are light years better! Thanks for posting and to you and Luis for hooking us up with the coupon. Super appreciated!
For anyone wondering, you can use your created Lora locally on your pc. i'm using it with flux nf4 on forge ui. but the results do seem a lot better running it on flux dev on replicate tho
You get better results because the first sentences of the prompt have more importance than the last ones. It works always like that! Basically I always try to use the prompt related to my lora as the first or last.
Man this is definitely one of your best videos, simple, straight to the point, and super impressive. Thumbs up. Edit: I wish I saw more examples at the end coz it's stunningly good (I don't maybe it's like "rewarding" more "fairly" after 20 minutes of watching, my brain needed more than these few examples 🙂🙂 )
Hello! Great video, I have a question. If I want to do several people, can I put all the images and train it or can it only be done with one person, since I tried it but it seems to mix the appearance of all the people into one.
Very cool. I’m amazed that you figured out this workflow… that is a lot of connecting the dots. I won’t do this one but I like the tutorials. Thanks Matt 😅
I followed the instructions and it worked perfectly. Only issue I have, is if there are more people in the generated image, they pretty much all look like me lol
Finally got it running with some tweaks mentioned in the comments. Unbelievable quality! I got the best results when I used the name of my AI alter ego often in the prompts
Very cool video - wanted to give it a go but it seems the Replicate Code has expired. Saw your pinned comment - sorry to make you do this again, but do you have another link? Keep up the good work in making clear and easy to follow videos like this
Hi, maybe someone knows what am i doing wrong. I tried filled grain and write too, in public and says to me this: Training failed. Command '['python', 'run.py', 'config/replicate.yml']' returned non-zero exit status 1.
Thank you for your guidance. However, after completing all the steps, when I run a prompt, the output image appears in a very short time (less than 2 seconds), then it immediately disappears, leaving only a placeholder image icon and the text "Output" instead. I always encounter this issue. How can I fix this issue? Thank you so much!
Hmmm.. I should know how to do this.. but how do I use my HuggingFace lora model in a local ComfyUI ? Do I name the lora.safetensors file a custom name and stick in in the models/loras folder? Gonna try that and find a flux workflow using a lora. YEP.. that worked in a Flux workflow. This was a fantastic video. Now I have to make several loras for characters and see if I can use em all in a generation. very nice. thanks.
great job matt, as usual. just a small request. would you mind switching your browser to Dark mode. i can hardly watch it's so bright and white now. thanks in advance
hello matt. i have a problem. as soon as i start to train it keeps failing. this is the error code. '['python', 'run.py', 'config/replicate.yml']' could you help?
On the video you said to choose the "public" option instead of "private" for the model. What happens if I choose the option of the model being private? I'm making a freelance AI animation project where it would be good if the people in the animation appear exactly as themselves on real life, but it is too risky to create a public model that anyone can access and generate any sort of images of these people, specially when one of them is very famous.
@@houmie very easy, In that repository you will find a setting button---> click on that ---> There must be a option change visibility---> Click on Private. That's it.
This is really cool, love this, thanks for walking us through it, just had a go and looks really good! Sweet! Also thanks for the credit allowing us to try!
Thank you so much for this video. It was very easy to follow along and get a working model. Cost me $3.10. It does realism very well but I'm struggling to make an illustrated version of myself. It seems like the Lora is squashing any stylization from the prompt. When I stylize with another Lora, the realism is still prominent in the image. Running locally in Comfy.
Thank you for the clear instructions as well as the $10 credit. Am I able to train multiple models on replicate, and then use them together to make an image with multiple trained subjects?
Update #2: I've just gotten word from Replicate that they have re-opened the $10 link! It should be working again. :)
Update #1: Apparently the link expired for Replicate. I was not told that the link would expire. I apologize for this. I honestly had no idea. I reached out to them and asked what happened and they told me that their finance team decided to turn it off because too many people were using it. I'm going to do what I can to rectify the situation, even if it means finding a different method people can follow to do it for free. Once again, I sincerely apologize. This decision was made my Replicate and they did not inform me that they would turn the link off if it got too much traffic.
yes doesnt work, maybe you will find other tools for free to show us
thanks for your dedication, a lot of can't afford to pay for gpu time and would really appreciate a free alternative to train flux lora, maybe kaggle perhaps
I’m a little late unfortunately, but thanks for letting us know. Love you videos ❤
Amazing thank you for your work!
@@mreflow Thank you sir 🔥🔥
all I want is to train locally, generate locally, and use a nice UI .
your GPU will get really hot so watch out, make sure you got an AC unit nearby
I would love to do it if I had a Nvidia H100 at home, or even an A6000, but no way my 2060S will handle anything worth the time
@@luizpaes7307 You can generate some pretty nice stuff with Stable Diffusion 1.5 on a 2060 with the right Checkpoint/Lora but yeah, not gonna work with Flux.
@@francisco444 I'll just wait for this winter to be sure :p
It's all been available for a long time; as long as your PC can handle it.
Hey Matt. We followed your instructions and generated some impressive images. Thanks for the detailed practical way you do the demos. Your dedication and passion are a cut above many.
Ok, there were a couple of things that I've learned from other vids that may have gotten missed with this one. 1. It must be set to public (which I think he was later addressed) and the other way I've learned and now do with no issues is 2. When creating your TOKEN you have three options - Fine-grained, Read, Write - you want to set it to WRITE and that solved all my problems. I hope it solves yours. Good Luck.
the fine grained setting is what got me, it must be to write
Can i make more than 1 lora with the same HF token?
Thank you so much! The error was driving me crazy. Unfortunately the results look nothing like me. It's probably my fault with the images I chose? There were I think 50. Maybe I should have prioritized the best ones.
Matt, this was a BRILLIANT Tutorial, worked like a charm on my first attempt !
Unfortunately I got a Python error when I tried. Pretty lame lol
@@brethnewsame
just watched this video, and the link is not working, unfortunately.
would u give us a new link ??
Your prompts are fabulous Matt keep it up!
That’s right man I love him
Link not working :(
Thank you so much Matt for looking out for us as always!
Regarding your question of positioning of the trigger word: LLMs in general pay more attention to the beginning and end of a prompt than they do to the middle (something which human brains *also* do, by the way 😁). So moving the trigger word to the beginning or end will indeed have a measurable effect on how often it includes the trained face/concept correctly.
For anyone getting the error Command '['python', 'run.py', 'config/replicate.yml']' returned non-zero exit status 1:
I also got the error, but it worked on the 2nd try with a couple of changes. Not sure what exactly was causing it but you can do all of these which will hopefully work for you.
1) Specify the model name instead of leaving it blank (black-forest-labs/FLUX.1-dev)
2) Rename all of your training images to be shorter (ex: person, instead of photo_of_person) (Edit: Probably not needed it's just something I randomly tried)
3) Make sure you request access to the black-forest-labs/FLUX.1-dev model on Hugging Face (I already did this prior, so probably not the issue, but who knows)
4) When creating your access token, scroll down to "Repositories permissions", select the location for your Lora, grant all permissions
It is likely step 4 that was the issue and what fixed it.
I would imagine first creating the repository and then the access token would also fix the issue.
i did 1 and 4 and it worked (had 3 done already). thank you
Its specifically step 3
@@avivolah9401 thanks!!!! i think this is it
I still got an error after doing all that, then I resized my images and there were two corrupt photos that I had to delete from my zip after taking claude's help. Now it is working fine.
OMG Matt. This training is blowing my mind! I'm so excited with the outputs I'm getting. And would have been crawling into a deep dark hole of despair trying to figure this out on my own. LOVE IT!
Hey Matt,
I just wanted to drop you a quick note to say how awesome I think your tutorials are. I’ve gone through the steps, and even though I haven’t had the chance to test yet, I can already tell that this is going to be a game changer for me.
It’s looking like we’re getting some rainy weather next week, so that’s when I’ll really dive into it. Just wanted to say thanks for making everything so easy to follow. You’re doing an amazing job, and I’m excited to see where this takes me./ Anna
There seems to be a giant issue in the script of the ai-toolkit that causes it to fail every time now.
It seems to be downloading EVERY image ever supplied to it succesfully. When you check the logs you can clearly see that it is downloading multiple sets of images with names of other people that (obviously) can't be found anymore or are so big in total size that the memory can't save them all locally anymore.
Seems like a 'little' oversight on the dev's part 😂
I get other people instead of me?
yeah that might be why i got my dog instead of myself sometimes.
Ohh that explains a lot. Its not really me sometimes I feel so nostalgic they look like family members I never met. Not much like me though
BRO!!! thank you ! your editing amazing love the new glitch style transition, and information presentation / lesson / 10 free , wow, this is one of your best videos yet. you the man !!
It worked like a charm.... but I want to use the model I created on hugging face on my local ComfyUI. Can you make a short tutorial on how to do this please!! Love your videos... keep up the good work!!
You download the safetensor from your Hugging face and include it in your lora folders. Onwards, you use it as any other lora. I didn't try yet, but that is what I will do when I will create mine.
Hi, the credit has expired, could you send a new one? Thank you!
They must have re-upped it.
coupon link has expird
A question - what happens if you don't want to upload the model to Hugging Face at all? Do you get a download link, or something else.
I read about this toolkit just yesterday but with the VRAM requirements it was a no go for me. Didn't know about Replicate but it has me interested and seems easy enough. Cheers.
Sir, for free credits it shows 'It looks like that invite has expired. Please ask whoever sent it to you to send a new one.'.
Can you please send the new invite link.
Thank you
bro i did everything you said and the training wont start
This is my first, formal AI training and I loved it. Thank you for your clear instructions
Please update the link
*Wow, I'm definitely gonna try this method! Great stuff Mr. 🐺*
It looks like that invite has expired. Please ask whoever sent it to you to send a new one.
You nailed it with your presentation; everything was spot on. Many thanx.
Thank you Matt for the clear instructions. I managed to follow along and create some amazing images.
Outputs are realistic af... Thanks Matt & Luca
i created my Lora & it went great thanks!
Can I use this using my GPU locally? Though either Comfy or Forge?
Thank you Matt. The replicate credit is expired, can you update it if possible?
Very good instructions I followed it to the letter and all went well. Thanks Matt !
Wow, thank you very much for providing the credits and your continuing great tutorials, really highly appreciated. I truly enjoy your content and am always thrilled when you upload something new. I especially love that you cover so many different AI topics and that you always take the time to teach your findings in a easy to follow way. Keep up the great work and thanks again for sharing.
Question, the $10 coupon no longer works... is there a page where I can upload or use my replicate lora? That is, use it and continue generating images on another page? (I have downloaded the lora replicate in a zip file).
A very nice and useful video! Thank you for that!
Can a model be retrained? Can I add a new set of photos to the model?
not working
Do the images I train the model on have to be a certain size?
Because I keep getting an error after I click on the train itself
I have the same problem.
I followed your instructions but after clicking an "create training" it always says: Command '['python', 'run.py', 'config/replicate.yml']' returned non-zero exit status 1. Cany ou help?
another problem might be that you have set the hugging face id wrong.
hi brother, i am facing the same issue, have you found the solution ?
@@Thecricketwatch you need to ask for the permission of the original model to use it (huggingface)
How can I keep my model private in HF and still using it in replicate?
Cannot understand how you only got 630K subscribers. By far, this is the best IA channel ever. Thks Matt
You haven’t watched enough
Please keep breaking the tools down like this I am taking notes here. Your tutorials have been invaluable and a new wave of awesomeness is clearly upon us.
Please update the link. it's not working
Definitely gonna try this out. Thanks for sharing everything in such detail man 🙌
Hey Matt, did you ever get Leonardo AI models to train properly on yourself? I AM STRUGGLING BIG TIMEEEE
Really appreciate that $10 replicate credit - immediate value and no fluff looks like the motto for this channel - i'm locked in
Super appreciate this. I followed your tutorial from last year, and these are light years better! Thanks for posting and to you and Luis for hooking us up with the coupon. Super appreciated!
For anyone wondering, you can use your created Lora locally on your pc. i'm using it with flux nf4 on forge ui. but the results do seem a lot better running it on flux dev on replicate tho
You get better results because the first sentences of the prompt have more importance than the last ones. It works always like that! Basically I always try to use the prompt related to my lora as the first or last.
Man this is definitely one of your best videos, simple, straight to the point, and super impressive. Thumbs up. Edit: I wish I saw more examples at the end coz it's stunningly good (I don't maybe it's like "rewarding" more "fairly" after 20 minutes of watching, my brain needed more than these few examples 🙂🙂 )
Is it possibile to use the trained Flux LoRA in automatic1111 or other sd based ui locally?
yes, I'm using it in DrawThings
Hello! Great video, I have a question. If I want to do several people, can I put all the images and train it or can it only be done with one person, since I tried it but it seems to mix the appearance of all the people into one.
I have the same question!
Very cool. I’m amazed that you figured out this workflow… that is a lot of connecting the dots. I won’t do this one but I like the tutorials. Thanks Matt 😅
Thanks Matt! It's finally getting easier and easier.
thank you thank you.... I'm in LOVE with your video....
Great, awesme instructions... Im gonna be taking notes....
man, you're AWESOME! thanks
I followed the instructions and it worked perfectly. Only issue I have, is if there are more people in the generated image, they pretty much all look like me lol
i faced this issue too - lmk if u found a solution>
🚫"You need to have a payment method set up to run this model"
Finally got it running with some tweaks mentioned in the comments. Unbelievable quality! I got the best results when I used the name of my AI alter ego often in the prompts
It worked perfectly, amazing pictures came out, very creative, thanks matt!
Fantastic video. Thanks for your effort in teaching people like me who have no idea how to code to use this magical tool.
great step by step instructions -do you have any for video?
Would be great to have a tutorial on using the generated LORA with ComfyUI
replicate inv is expired
when will be the link working again? please answer
AWESOME video Matt! You're the GOD of AI! Thank you for sharing your insights. Just tried everything you taught here and enjoyed it a lot!
This is super good. The coupon code isnt working :-(
didnt work for me, displays errors
unfortunately the link is already cant be used anymore
Thanks for sharing, but the AI training mode is no longer available under Lucataco account, what should I do? please advise. Thanks
Tutorial videos are always welcome. It’s how us little folk get a leg up on the rich and powerful. 😊
"It’s how us little folk get a leg up on the rich and powerful" .... - Dream on
Great video!! Is this same workflow capable of training and producing product photos. Like shirts and glasses?
I really liked the way you presented it.
Link no good anymore.. ;(
Following your instructions, I successfully created AI images of myself. It was such fun. Thank you very much!
Very cool video - wanted to give it a go but it seems the Replicate Code has expired. Saw your pinned comment - sorry to make you do this again, but do you have another link? Keep up the good work in making clear and easy to follow videos like this
Great tutorial Matt! Props for the credits you landed for fans. I do enjoy Replicate and playing with all the models with Flux indeed being the best.
Hi, maybe someone knows what am i doing wrong. I tried filled grain and write too, in public and says to me this: Training failed.
Command '['python', 'run.py', 'config/replicate.yml']' returned non-zero exit status 1.
I also got the same error message
Super! Muchas gracias Matt, sería genial si tuvieras subtítulos en español tus videos.. Saludos y muchas gracias por tu aporte.
Skintexture is too smooth and skin too glossy/shiny. How can i make it more realistic?
⚠ It looks like that invite has already been used. Please ask whoever sent it to you to send a new one.
This was AWESOME, Matt! Love all your videos and how you break it all down. Keep going! :)
Free Replicate credits are not woking its says its expired can you pleas provide a new one?
You rock man thank you so much! I'm using these pics for my IT resume. Appreciate ya.
Question regarding image resolution...can the model give us higher resolution images? On 1:1 ratio I'm getting 1024x1024.
This is literally one of the most valuable YT vids ever put out there. Holy f*ck, does it work!
Amazing tutorial! Thank you so much. What if I want to fine tune another person face on to the fine tuned model with my face
update the link bruh
Thank you for your guidance.
However, after completing all the steps, when I run a prompt, the output image appears in a very short time (less than 2 seconds), then it immediately disappears, leaving only a placeholder image icon and the text "Output" instead.
I always encounter this issue. How can I fix this issue?
Thank you so much!
Hmmm.. I should know how to do this.. but how do I use my HuggingFace lora model in a local ComfyUI ? Do I name the lora.safetensors file a custom name and stick in in the models/loras folder? Gonna try that and find a flux workflow using a lora. YEP.. that worked in a Flux workflow. This was a fantastic video. Now I have to make several loras for characters and see if I can use em all in a generation. very nice. thanks.
Thanks for the voucher. Trying to follow your instructions now.
great job matt, as usual. just a small request. would you mind switching your browser to Dark mode. i can hardly watch it's so bright and white now. thanks in advance
hello matt. i have a problem. as soon as i start to train it keeps failing. this is the error code. '['python', 'run.py', 'config/replicate.yml']' could you help?
Hi Matt,
Thank you for this. It's exactly what I was looking for.
On the video you said to choose the "public" option instead of "private" for the model. What happens if I choose the option of the model being private?
I'm making a freelance AI animation project where it would be good if the people in the animation appear exactly as themselves on real life, but it is too risky to create a public model that anyone can access and generate any sort of images of these people, specially when one of them is very famous.
Did you find an answer for that? I have the same question
@@AndresArosemena Click on 'Settings' in the created repository. There, you will find an option to change the model's visibility. Set it to 'Private'.
That's a good question, have you been able to solve it? Otherwise I find this to be quite risky from a privacy point of view.
@@houmie very easy, In that repository you will find a setting button---> click on that ---> There must be a option change visibility---> Click on Private. That's it.
“A lot of time has passed”. No it hasn’t. That’s the amazing thing.
This is really cool, love this, thanks for walking us through it, just had a go and looks really good! Sweet! Also thanks for the credit allowing us to try!
hello bro your contents are all great please update the replicate invite link 🥺
Yeah, sadly already expired @mreflow, wanna hook us up with another link?
Hi Matt, the link to the $10 credit coupon has expired. Can you please ask Luctaco if he can renew it? Cheers Aldo
Thanks so much. I learn so much each time I watch your brilliant videos. You are my real-life super hero. 🎉
Thank you so much for this video. It was very easy to follow along and get a working model. Cost me $3.10. It does realism very well but I'm struggling to make an illustrated version of myself. It seems like the Lora is squashing any stylization from the prompt. When I stylize with another Lora, the realism is still prominent in the image. Running locally in Comfy.
can i download the lora model to use in comfy ui?
waiting for that answer too, no luck yet.
Me too
yes, i download it, all works, but coupon expired
I have about 50 images. Is there a benefit or downside to using more images for the training?
Thank you for the clear instructions as well as the $10 credit.
Am I able to train multiple models on replicate, and then use them together to make an image with multiple trained subjects?